The market
never sleeps.
You do.

Two opposing desires.

When you trade, you want two things at once: to win big and to lose little. The problem is that the two work against each other: every time you reach for more gain, risk climbs.

The chart on the right illustrates this. The vertical axis is gain. The horizontal axis is the worst possible drop (what we call risk).

A thousand possible strategies.

Each little dot that appears is a tested trading strategy: a different setting, a different behavior. Botlyz regularly tests more than a thousand.

You can see they form a cloud: each one has its own gain and its own risk.

The "top-left"
is ideal.

In the top-left corner: what we dream of. Lots of gain, little risk. That is where we want to be.

In the bottom-right corner: disaster. Little gain, lots of risk. And no one can win without taking on some risk.

The frontier of the best.

NSGA-II draws a curved line: the best possible performance for each accepted level of risk. This line is called "the Pareto front".

The green dots sit on it: these are the unbeatable strategies. The others, further back, are beaten by at least one green strategy. So we discard them.

The final choice.

What remains is choosing which green dot on this frontier. Botlyz uses a third criterion: stability over time. Not the highest gain (often unstable), not the lowest risk (often lazy). The point that holds up over the long run.

The problem:
too many settings.

A trading strategy has plenty of parameters to set: when to buy, when to sell, how much to risk, on which pair… Combine them all and the number of possibilities explodes.

First attempt:
at random.

We start by trying about thirty random settings. On the chart, each dot = one tested setting. The higher the dot, the better the setting works.

The result looks like a cloud: some settings are good, others bad. But at this stage, we still don't know why.

The TPE trick.

TPE (Tree-structured Parzen Estimator) is a statistical algorithm that looks at the attempts already made and guesses the shape of the terrain.

The green curve is that guess. The peaks = the zones where it works well. The troughs = the zones to avoid. It is exactly like a treasure map drawn from clues.

Now we dig in the right spot.

TPE focuses its new attempts around the peaks on its map. With each new attempt, it refines its guess. The more it makes, the sharper the map gets.

Instead of searching an entire house, it digs where it heard a noise. Logical, and fiercely efficient.

×12 faster.

Concrete result: with 3,200 well-placed attempts, TPE finds the same optima as a brute-force search that would require more than 40,000.

The "rote memory" trap.

A student who learns the answers by heart will score 20/20 on the exam they know, and 0/20 on the next. They understood nothing, they just memorized.

For a trading algorithm, the trap is exactly the same: it can "learn" the past so well that it becomes unable to trade the future. The technical term: "overfitting".

The countermeasure:
hide a portion.

Here is our market history: 360 days of real prices. The solution is simple: instead of showing the algorithm everything, we hide a portion from it.

It trains only on the light window on the left (60 days). The rest, it never sees.

Then we give it the exam.

Once trained, we show it the green window right after (15 days), the one it has never seen. And we check whether its trades are good or bad.

If the results stay good on unseen data, it means it did not memorize: it genuinely understood the logic.

We repeat 5 times.

A single passed exam can be luck. So we shift the windows through time and start over. 5 different exams, over 5 different periods.

The algorithm has to pass every exam, not just the average. Otherwise, we reject it.

Validated.
Or rejected.

There is the filter. Out of 100 strategies that look good in training, only about 13 pass all 5 exams. The other 87, we discard.

Three traps people forget.

An "idealistic" backtest assumes you can buy at exactly the displayed price, in unlimited quantity, and without paying any fees. All three assumptions are false in real life.

Ignore them, and the displayed return is flattered. And disappointment is inevitable.

Trap #1: slippage.

You see the price at 3,412 €, you decide to buy. By the time the order reaches the market (a few milliseconds), the price has moved to 3,414 €.

That is slippage: the difference between the price you saw and the price you got. Often unfavorable.

Trap #2: liquidity.

On the market there is always someone on the other side. But on some less popular pairs, the order book is thin. If you buy big, you drain the book and the price climbs as you go.

The consequence: on a medium-sized order, the average purchase price can be noticeably higher than the displayed price.

Trap #3: fees.

Every conventional exchange charges 0.1% per trade (taker). Across the ~370 annual trades of a strategy like SIGMA AWF, that is 30% of capital eaten in fees before you even talk about PnL. Most "profitable on paper" strategies stop being profitable once those fees are counted properly.

Botlyz runs on Lighter (0% maker/taker) and instead applies 0.01% to 0.10% per trade on-chain, depending on the strategy. Overall, the cost stays on par with a conventional exchange, but you also get the turnkey strategy, the routing, and real-time monitoring. Everything is included, without exception, in our simulations: the displayed performance is net of fees.

The result:
the truth.

When you see a return on a Botlyz strategy, it is what it would have actually earned, with slippage, liquidity and fees already deducted. Not a "magic" return that evaporates in production.

Sharpe 3.71.

The Sharpe ratio is the universal score of a strategy: how much you earn for each unit of risk taken. The higher, the better.

Benchmarks: above 1 = decent strategy. above 1.5 = good. above 2 = excellent.

Sortino 5.60.

Sortino is Sharpe in its improved form. It looks only at losses, ignoring sharp upswings. Makes sense: when you're winning, who cares about volatility.

Our 5.60 means this: for 1 € of potential losses, the strategy captures 5.60 € of gains. The ratio is heavily positive.

Drawdown −10.1%.

The "max drawdown" is the worst observed drop in capital. From the highest peak down to the lowest trough before recovering.

Over 1 year of OOS backtest, the worst moment was −10.1% (spread over 24 days). An active strategy stays acceptable below −20%.

Calmar 22.0.

Calmar combines two things: the annual return divided by the worst drawdown. It is a measure of "peace of mind": how much you earn per unit of potential pain.

A Calmar of 22.0 is exceptional. Above 3, it is already considered that the return amply justifies the drawdown risk, and here we are well beyond that.

Win Rate 60.7%.

Across 367 trades, nearly 6 out of 10 ended in profit. That is well above the average for active trading (often 45 to 55%).

And the Profit Factor of 1.93 completes the picture: for every 1 € lost, the strategy earns 1.93 €. Gains comfortably outweigh losses.

Validated for production.

All these scores were computed on the last 12 months of data, which the strategy never saw during training. Had it underperformed over that period, we would have rejected it.